Homemade Pasta · Research
Google's AI Overviews reduce organic clicks by roughly 40% on the queries where they appear. Optimizing for citation recovers part of that. A ceiling of 38% does not come back at all. This is what the evidence supports, what it is worth, and what to put on the board first.
Why this exists
Homemade Pasta runs paid media for clients in categories with little in common. Organic search volume is down across the client base. Their organic teams keep asking the same question the rest of the industry is asking, and the answers on offer are a single vendor study or a single account’s dashboard.
Neither one plans a year. Published estimates of the AI Overview effect range from 15% to 58%, a band too wide to budget from. One account cannot narrow it either, because a single set of dashboards cannot separate an AI Overview effect from a ranking change, a seasonal shift or a tracking break.
So we read the evidence ourselves. This page consolidates what holds up into one calibrated number, then turns that number into a plan. What to measure first, what to fund, what to stop funding, and what to take out of the forecast permanently.
Executive summary
Three findings determine how much to invest, what to forecast, and how to split search budget.
The practical consequence is a change in forecast, not only a change in tactics. An organic plan built on recovering pre-AI-Overview volume will miss. An organic plan built on capturing the largest available share of a smaller pool, with search budget rebalanced toward the channel that holds up better, will not.
Recommendations
| Action | Why it comes first | Quantified impact |
|---|---|---|
| 1. Measure exposure before anything else | The share of organic impressions sitting on AI Overview-present queries determines the size of every other number here. Most organizations do not have it. | Below 15% exposure, this is a minor program. Above 40% it is the highest-value organic work available. |
| 2. Audit crawler access and snippet eligibility | The only work every major platform documents as necessary. A nosnippet directive or a bot rule at the CDN removes AI Overview eligibility outright, and neither is visible in a standard SEO audit. | Removes a binary risk of total exclusion. Costs days, not quarters. |
| 3. Fund citation work in proportion to exposure | Ranking still governs eligibility. It no longer governs the outcome, so citation is a separate objective from position. | Doubling a citation rate from 25% to 50% returns 7% to 23% more organic clicks on the affected query set. |
| 4. Hold or increase paid coverage on affected clusters | Informational and comparison queries carry the heaviest AI Overview exposure. Paid performance on those queries has been stable while organic has compressed. | Paid retains 80% of its clean-page click-through where the brand is cited. Organic retains 62%. |
| 5. Reset the organic forecast | The structural shortfall is not addressable by content, markup or technical work. Carrying it as a performance gap misallocates effort for several quarters. | Perfect citation performance returns 62% of pre-AI-Overview clicks on affected queries. The remaining 38% should be planned as permanent. |
Part One
The size of the loss, where it concentrates, and what happens to paid search on the same pass.
1.1 The loss
Estimates of the organic click decline when an AI Overview is present range from 15% to 58%. Four are observational. One randomized users, and it lands in the middle of the range at 39.8%. That figure should anchor any forecast.
The spread reflects measurement populations rather than disagreement about direction. The narrowest estimate was taken only on keywords where the site already held a ranking. The widest spans two years, is position-one specific, and is desktop only.
Participants were randomly assigned through a browser extension to standard Google Search, to a version with AI Overviews suppressed, or to AI Mode. Approximately 1,065 US desktop Chrome users and 68,089 searches.
Conditional on an AI Overview appearing, outbound organic clicks fall 39.8% and zero-click searches rise 34.5%. Sponsored clicks do not move. Total search frequency does not move.
A secondary result carries equal weight for planning. Google's stated explanation is that AI Overviews suppress low-value clicks and that surviving clicks are higher quality. Measured against the control group, downstream engagement per click did not improve.5
A separate behavioral panel establishes the corollary. Users click a link inside an AI Overview in approximately 1% of visits to pages carrying one, against 8% clicking a traditional organic result on those pages and 15% on pages with no AI summary.9 Citation is not a distribution channel. It is a position advantage on the organic clicks that still occur.
Query shape governs exposure more than industry does. Across 55,393 queries, question-form queries triggered an AI Overview 64.7% of the time against 9.5% for non-question queries, a 6.8-fold difference. “How” queries reached 84.3%. Single-word queries reached 9.9%. A separate dataset places comparison queries, the “X versus Y” format, at 95.4%.1, 2
These groups overlap rather than partition. The sample is of trending queries, which skew question-shaped, so the direction is firm and the absolute rates are probably high. Preprint, under review.1
By intent, informational queries carry most of the exposure at 36% prevalence, against 8% commercial and 5% transactional. Transactional queries are now affected as well: organic click-through on AI Overview-present, uncited transactional queries fell from 4.17% to 2.15% across 2025.2
Brand strength is protective. Branded keywords with an AI Overview present gained 18.68% click-through in one study while non-branded lost 19.98%, and only 4.79% of branded keywords triggered an AI Overview at all.3 Across thousands of publisher sites, search referral traffic over two years fell 60% for small publishers, 47% for mid-sized and 22% for large.4
Google documents the architecture, and it determines which levers exist. AI Overviews are produced by retrieval-augmented generation, described by Google as a technique “used to improve the quality, accuracy, and freshness of AI responses by relying on our core Search ranking systems to retrieve relevant, up-to-date web pages from our Search index.” The same systems that rank organic results supply the candidate set, which is why ranking remains the strongest available lever.
Google also generates fan-out queries, a set of concurrent related searches run alongside the query the user typed. A page can therefore be cited without ranking for the visible query.
Independent measurement confirms the second mechanism is material. Across 55,393 queries, 29.8% of AI Overview-cited domains did not appear anywhere in the first-page organic results displayed beside them. Source selection is partly a separate retrieval process rather than a re-ranking of page one.1
Two further findings from the same study bear on content strategy. AI Overviews cite more credible sources than the organic results beside them, scoring 0.732 against 0.645 on a continuous credibility measure with a 95% confidence interval of 0.085 to 0.089 on the gap. And they under-cite user-generated content, which accounts for 14.2% of AI Overview references against 41.4% of first-page results.
1.2 Paid search
Both channels compress when an AI Overview appears. They do not compress equally, and the gap is wide enough to change how a search budget is allocated.
Where the brand is present but uncited, organic falls 72% and paid 43% on the figures as published. Section 2.1 explains why the uncited organic figure is contaminated; on the publisher's adjusted number the organic fall is 52%.
The trend reinforces the level. Across all query types, paid click-through on AI Overview queries rose from 14.6% in January 2025 to 16.2% in February 2026, while paid click-through on queries with no AI Overview declined from 26% to 21.8%. Whatever is compressing paid performance on clean results pages, AI Overviews are not the cause.
Allocation consequence. On informational and comparison clusters, paid is currently the more predictable of the two channels. Treating the organic decline as a reason to reduce total search investment moves spend away from the half of the page that is holding up.
Part Two
The value of a citation, the ceiling above it, and a model that portions both against a specific set of impressions.
2.2 Sizing
The figures above are expressed per million impressions. The number that determines their significance is the share of a given site's organic impressions that sit on AI Overview-present queries.
Value of a citation program
Informational queries with an AI Overview present. All outputs recalculate, including the counterfactual click volume at the no-AI-Overview rate.
A defined program with a named owner. The gain sits inside what a quarter of focused work can move, though it will not appear cleanly in Search Console reporting.
Uses 2.07% cited and 3.35% no-AI-Overview CTR throughout, with the uncited rate set by the toggle. The 3.35% is a contemporaneous rate on queries that do not trigger an AI Overview, not a measured pre-2024 baseline, and those are different query populations. Informational intent only, from a single 53-brand dataset whose publisher states it cannot claim causation. Treat the output as an order of magnitude, not a forecast.
Part Three
The practices the evidence supports, the practices it does not, and the one that burns the dish.
3.1 Practices
Ordered by the strength of the evidence behind each.
90% of AI Overviews cite at least one top-10 organic result, and a page ranking first is cited about 43% of the time, falling to 7% at position 20. Only 38% of individual citations are top-10 URLs, down from 76% a year earlier. Ranking governs eligibility rather than outcome.8
Google requires a page to be indexed and eligible to display with a snippet in order to appear as a supporting link. A nosnippet or max-snippet:0 directive removes AI Overview eligibility outright. This is the most common self-inflicted exclusion and it costs nothing to verify.
Google's published checklist specifies that crawling be allowed “in robots.txt, and by any CDN or hosting infrastructure”. Firewall and bot-management rules do not appear in a robots.txt audit and are the dominant real-world cause of exclusion. Google-Extended governs Gemini training and grounding only and has no bearing on AI Overviews.
Google's guidance draws the boundary at content that “could easily be produced by a generative AI model”. With a median of eight citations per AI Overview and a measured preference for higher-credibility sources, undifferentiated content competes for a materially narrower opening than it did against ten ranked links.
Analysis of Google's grounding layer found a median budget of 1,929 words per query and 377 words selected per source, with coverage falling from 61% on pages under 1,000 words to 13% on pages over 3,000. The measurement covers the Gemini grounding API rather than AI Overviews directly and dates from December 2024, so it is indicative rather than definitive.
1,885 pages that added JSON-LD, matched against 4,000 control URLs on other domains in a difference-in-differences design with an event study confirming no pre-treatment drift, produced a 4.6% decline in AI Overview citations, statistically significant. Cited pages are three times more likely to carry schema, and that correlation did not survive causal testing. Schema remains justified on rich results, which is a documented and measurable benefit.
Google states that Google Search ignores the file. Across 137,210 domains, 97% of published llms.txt files received zero requests over a month, and the largest single category of requester was SEO audit tooling.
Google states there is “no requirement to break your content into tiny pieces for AI to better understand it” and that “there's no ideal page length”. Microsoft's guidance for Copilot takes the opposite position, making this a per-platform decision rather than a general practice.
AI Overviews under-cite user-generated content relative to the organic results beside them, at 14.2% of references against 41.4% of first-page results. Source-type share is also unstable: Reddit appeared in roughly 60% of ChatGPT responses before mid-September 2025 and roughly 10% after, with no content changes anywhere.
Google deprecated the FAQ rich result in May 2026. The feature that markup powered no longer appears in search results.
Across 107,352 pages appearing in AI Overviews and AI Mode, Largest Contentful Paint correlated with AI visibility at −0.12 to −0.18 and Cumulative Layout Shift at −0.05 to −0.09. Performance is a threshold to clear rather than a lever to pull.
Google's guidance states that creating separate content for query variations, “for example, by focusing on other queries that people have asked, or fan-out queries”, primarily to manipulate rankings or generative AI responses, “violates Google's scaled content abuse spam policy”. Google extended its spam policies in May 2026 to cover manipulation of generative AI responses explicitly.
Appendix
Supporting detail behind the findings above. Open what is relevant.
A.1 Measurement
This constrains every claim an organization can make about its own AI Overview performance, and every claim made on its behalf by a vendor. Google's Search Console generative AI report provides impressions only, with no clicks, no click-through rate, no position and no query dimension. AI Overview clicks are folded into the standard performance report with no means of segmenting them. Both limitations are documented by Google rather than inferred.
The practical consequence is that a citation program must be measured on citation rate, which is observable through third-party rank tracking, rather than on attributable clicks, which are not.
| Source | Provides | Does not |
|---|---|---|
| Search Console, generative AI report | Impressions from AI Overviews and AI Mode by page, country, device and date. Worldwide since 31 August 2026. | No clicks, click-through rate, position or query dimension. 1,000-row limit, manual export, not exposed in the documented API. Data begins 18 May 2026. A subset of Web-type impressions, so the two are never summed. |
| Search Console, standard report | AI Overview clicks are included in the totals. | Undifferentiated. An AI Overview occupies a single position with all links assigned that position, which distorts average position. |
| Bing Webmaster Tools, AI Performance | The only first-party query-side citation data available: grounding query intents, topic clusters, citation share. | No API. Microsoft states citation share does not represent traffic share or assign quality scores. |
| GA4, AI Assistant channel | Assistant referrals, from 13 May 2026. | Not retroactive. Perplexity is excluded. Referrer-less traffic remains in Direct. |
Google formerly appended a text fragment to some AI Overview citation clicks, and a technique for capturing it in a tag manager circulated widely. Two independent practitioners who published that technique have marked it discontinued, reporting that Google moved to inline links without text fragments around May 2026. The frequently quoted finding that 22.4% of AI Overview traffic misattributes to Direct derives from this method and is now historical.
A custom channel group, configured alongside the native channel rather than in place of it. Condition on Session source matches regex, ordered above Referral, since GA4 evaluates channels in sequence. Add a second condition group where Source matches the same pattern and Medium is (not set), to capture sessions where a source-only parameter prevents referrer fallback. GA4 regex matching is case-sensitive.
^(?:chatgpt\.com|chat\.openai\.com|claude\.ai|perplexity\.ai|gemini\.google\.com|copilot\.microsoft\.com|meta\.ai|grok\.com|deepseek\.com|you\.com)$
This captures the assistants. It cannot capture AI Overviews or AI Mode, which arrive as google / organic and are indistinguishable from a ranked-result click.
A falling click-through rate is frequently not lost clicks. In one 53-brand dataset, click-through fell 52% in a single month while absolute clicks rose from 398,798 to 400,271. Impressions doubled. The following month is when clicks actually declined, by approximately 25%.
Click-through rate should never be reported without absolute clicks alongside it. Where AI surfaces inflate impressions, the ratio alone is uninformative.
Per-query citation tracking has a noise floor above the effect size. Across 43,000 keywords each with at least sixteen recorded AI Overviews, consecutive AI Overviews for the same query shared only 54.5% of their cited URLs. An AI Overview has a 70% chance of changing between observations and persists about 2.15 days before substantive rewriting. Semantic similarity between consecutive responses is 0.95, so meaning is stable while citations churn.12
A citation KPI therefore requires a rolling aggregate across a large, fixed query set, a window of four weeks or more, a stated band rather than a point estimate, and a query set held constant between periods.
Google's Search Console control that excludes a property from AI Overviews appears to offer a deliberate holdout. Three problems make it unsuitable on a revenue-bearing property. Recovery time is undocumented: Google publishes only that exclusion applies within one to two days and states nothing about re-inclusion. Granularity is property-level, producing one treated unit and one control unit that differ systematically. And the outcome is confounded, because Search Console reports no AI-feature clicks, so the measurable quantity is total Web clicks. Google states that sites opting out “will not receive any traffic or impressions from these features”.
A page-group holdout using nosnippet provides per-URL granularity but also removes the ordinary organic snippet, so the measured effect combines both. Practitioners have documented content reappearing days later despite the directive remaining in place.
The defensible design is a pre-and-post on aggregate citation rate across a fixed query set, with a matched control set of queries that do not trigger AI Overviews. It does not isolate causality but it separates site-level movement from platform-level movement. A matched-market test on content built for high-prevalence clusters, measuring branded search volume and direct sessions rather than organic clicks, is the only design that would establish whether citation carries value beyond the click. No such test has been published.
A.2 Adjacent channels
ChatGPT and the other assistants generate high-intent traffic at a volume roughly two orders of magnitude below the AI Overview question. Tripling assistant referrals from 0.3% adds about six tenths of a percentage point to total sessions. Where 30% of organic impressions sit on AI Overview-present queries, AI Overviews are already moving a double-digit share of organic volume. That comparison determines sequencing.
Growth is normalizing. Adobe measured the same US retail series at +4,700% year on year in July 2025, against a near-zero base, and at +62% in July 2026. Those are separate reports against incomparable bases, which is why four-figure growth rates should not be quoted as current.10
The traffic performs well on arrival. In US retail, AI-referred visitors convert at a 60% higher rate than non-AI traffic and generate 53% more revenue per visit, reversing a substantial deficit measured in late 2024. Two qualifications apply: the traffic is heavily self-selected, and a study of roughly 82,000 websites in mid-2025 found the opposite on engagement, most plausibly explained by the twelve-month gap and by sector mix.
Across 127,198 citations from approximately 16,400 AI answers spanning five engines, only 2.7% of cited domains were cited by all five, and 69.6% were cited by exactly one.11 No single set of changes produces visibility across all of them.
| Dimension | Google AI Overviews | Assistants |
|---|---|---|
| JavaScript | Googlebot renders it. | Measured crawler behavior found OAI-SearchBot, GPTBot, ClaudeBot and PerplexityBot do not execute JavaScript. Measured December 2024, with no published replication since. |
| Freshness | Cited pages average 1,432 days old, 16 days older than the organic results beside them. Not a lever. | ChatGPT cited pages average 958 days, roughly 458 days newer than organic. A lever. |
| Sources per answer | Median 8 references. | ChatGPT averages about 15. Gemini averages 3. |
| Freshness submission | Google is not listed among IndexNow participants. | IndexNow reaches Bing and therefore Copilot. Not OpenAI, Anthropic or Perplexity. |
Sequencing. The access and snippet audit first, since it is inexpensive and covers both surfaces. The AI Overview citation program second, sized against measured exposure. Assistant-specific work third, and revisited if assistant referral share on a given site passes roughly 2%.
A.3 Sources
Numbered references appear inline throughout. Each entry states the publication date, the sample, and whether the publisher sells a product in the category the finding supports.